Noise Suppression RNN on Intermittent Device

Abstract

The field of edge ML has been evolving rapidly in the recent years. Understanding the various levels involved in deploying a ML model on an edge device like micro controller with limited resources will a challenging task. There are various stages involved in this process and a deep understanding of each field is required. In this context this will be a report on one such deployment of an advanced recurrent neural network which has never been deployed on a 16 bit microcontroller before(to the best of my knowledge).

This report will discuss each level of deployment in detail, gives the necessary background information. One other important aspect of this report is the survey of various applications in tiny ML and benchmarks used. The challenges faced could be taken up by budding researchers for future work. This report will show the results obtained from this deployment, study the effects of quantization and how the accuracy has been affected.

Implementing this model an intermittent system will also be one of the goals of this project as an rnn has never been implemented on BOBBER(Intermittent System) before and this throw some light on the working of such a complex model with intermittent power and volatile memory. One more step in this project is to assess the dependency level of the RNN on the previous outputs, based on which its deployment requirements like memory will be estimated for the intermittent system. The future works that can be performed on this model will also be discussed along with the results and conclusion.